Segment Anything Model (SAM)
The Segment Anything Model (SAM) from Meta introduced promptable image segmentation: given a point, box, or mask prompt, it returns a mask for the indicated object, with a lightweight decoder fast enough for interactive use.
Introduction to SAM: The Segment Anything Model#
The Segment Anything Model, or SAM, is an image segmentation model built for promptable segmentation across a wide range of image analysis tasks. SAM is the core of Meta's Segment Anything project, which introduced a new model, task, and dataset for image segmentation.
SAM's design allows it to adapt to new image distributions and tasks without prior knowledge, a capability known as zero-shot transfer. Trained on the SA-1B dataset, which contains more than 1 billion masks across 11 million curated images, SAM shows strong zero-shot performance that is often competitive with or better than prior fully supervised results.
SA-1B example images. Images with overlaid masks from the SA-1B dataset. SA-1B contains 11M diverse, high-resolution, licensed, and privacy-protecting images and 1.1B high-quality segmentation masks. These masks were annotated fully automatically by SAM, and as verified by human ratings and numerous experiments, are of high quality and diversity. Images are grouped by number of masks per image for visualization (there are ∼100 masks per image on average).
Key Features of the Segment Anything Model (SAM)#
- Promptable Segmentation Task: SAM was designed with a promptable segmentation task in mind, allowing it to generate valid segmentation masks from any given prompt, such as spatial or text clues identifying an object.
- Advanced Architecture: The Segment Anything Model uses a heavyweight image encoder, a prompt encoder, and a lightweight mask decoder. This architecture enables flexible prompting, real-time mask computation, and ambiguity awareness in segmentation tasks.
- The SA-1B Dataset: Introduced by the Segment Anything project, the SA-1B dataset features over 1 billion masks on 11 million images. As the largest segmentation dataset at its release, it provides SAM with a diverse and large-scale training data source.
- Zero-Shot Performance: SAM displays strong zero-shot performance across various segmentation tasks, making it a ready-to-use tool for diverse applications with minimal need for prompt engineering.
For an in-depth look at the Segment Anything Model and the SA-1B dataset, please visit the Segment Anything GitHub and check out the research paper Segment Anything.
SAM-family models (SAM 2.1, SAM 3, and SAM 3.1) power the smart annotation feature on Ultralytics Platform, enabling click-based masking for fast dataset labeling.
Available Models, Supported Tasks, and Operating Modes#
This table presents the available models with their specific pretrained weights, the tasks they support, and their compatibility with different operating modes like Training, Validation, Inference, and Export. Supported modes are indicated by ✅ and unsupported modes by ❌.
| Model Type | Pretrained Weights | Tasks Supported | Training | Validation | Inference | Export |
|---|---|---|---|---|---|---|
| SAM base | sam_b.pt | Instance Segmentation | ❌ | ❌ | ✅ | ❌ |
| SAM large | sam_l.pt | Instance Segmentation | ❌ | ❌ | ✅ | ❌ |
How to Use SAM: Versatility and Power in Image Segmentation#
The Segment Anything Model can be used for many downstream tasks beyond its training data, including edge detection, object proposal generation, instance segmentation, and preliminary text-to-mask prediction. With prompt engineering, SAM can adapt to new tasks and data distributions in a zero-shot manner.
SAM prediction example#
Segment image with given prompts.
from ultralytics import SAM
# Load a model
model = SAM("sam_b.pt")
# Display model information (optional)
model.info()
# Run inference with bboxes prompt
results = model("ultralytics/assets/zidane.jpg", bboxes=[439, 437, 524, 709])
# Run inference with single point
results = model(points=[900, 370], labels=[1])
# Run inference with multiple points
results = model(points=[[400, 370], [900, 370]], labels=[1, 1])
# Run inference with multiple points prompt per object
results = model(points=[[[400, 370], [900, 370]]], labels=[[1, 1]])
# Run inference with negative points prompt
results = model(points=[[[400, 370], [900, 370]]], labels=[[1, 0]])Segment the whole image.
from ultralytics import SAM
# Load a model
model = SAM("sam_b.pt")
# Display model information (optional)
model.info()
# Run inference
model("path/to/image.jpg")- SAM segments the whole image when you don't pass any prompts (bboxes/points/masks).
Set the image once, then run prompt inference multiple times without re-running the image encoder.
import cv2
from ultralytics.models.sam import Predictor as SAMPredictor
# Create SAMPredictor
overrides = {"conf": 0.25, "task": "segment", "mode": "predict", "imgsz": 1024, "model": "mobile_sam.pt"}
predictor = SAMPredictor(overrides=overrides)
# Set image
predictor.set_image("ultralytics/assets/zidane.jpg") # set with image file
predictor.set_image(cv2.imread("ultralytics/assets/zidane.jpg")) # set with np.ndarray
results = predictor(bboxes=[439, 437, 524, 709])
# Run inference with single point prompt
results = predictor(points=[900, 370], labels=[1])
# Run inference with multiple points prompt
results = predictor(points=[[400, 370], [900, 370]], labels=[1, 1])
# Run inference with negative points prompt
results = predictor(points=[[[400, 370], [900, 370]]], labels=[[1, 0]])
# Reset image
predictor.reset_image()Segment everything with additional args.
from ultralytics.models.sam import Predictor as SAMPredictor
# Create SAMPredictor
overrides = {"conf": 0.25, "task": "segment", "mode": "predict", "imgsz": 1024, "model": "mobile_sam.pt"}
predictor = SAMPredictor(overrides=overrides)
# Segment with additional args
results = predictor(source="ultralytics/assets/zidane.jpg", crop_n_layers=1, points_stride=64, min_mask_region_area=100)All the returned results in the above examples are Results objects which provide access to the predicted masks and the source image.
- For more
Segment everythingargs, see thePredictor/generateReference.
SAM Comparison vs YOLO#
Here we compare Meta's SAM-b model with Ultralytics segmentation models, including YOLO26:
| Model | Size (MB) | Parameters (M) | Speed (CPU) (ms/im) |
|---|---|---|---|
| Meta SAM-b | 375 | 93.7 | 41703 |
| MobileSAM | 40.7 | 10.1 | 23802 |
| FastSAM-s with YOLOv8 backbone | 23.9 | 11.8 | 58.0 |
| Ultralytics YOLOv8n-seg | 7.1 (52.8x smaller) | 3.4 (27.6x less) | 24.8 (1682x faster) |
| Ultralytics YOLO11n-seg | 6.2 (60.5x smaller) | 2.9 (32.3x less) | 24.3 (1716x faster) |
| Ultralytics YOLO26n-seg | 6.7 (56.0x smaller) | 2.7 (34.7x less) | 25.2 (1655x faster) |
This comparison demonstrates the substantial differences in model sizes and speeds between SAM variants and YOLO segmentation models. While SAM provides unique automatic segmentation capabilities, YOLO models, particularly YOLOv8n-seg, YOLO11n-seg and YOLO26n-seg, are significantly smaller, faster, and more computationally efficient.
SAM speeds measured with PyTorch, YOLO speeds measured with ONNX Runtime. Tests run on a 2025 Apple M4 Air with 16GB of RAM using torch==2.10.0, ultralytics==8.4.31, and onnxruntime==1.24.4. To reproduce this test:
from ultralytics import ASSETS, SAM, YOLO, FastSAM
# Profile SAM-b, MobileSAM
for file in ["sam_b.pt", "mobile_sam.pt"]:
model = SAM(file)
model.info()
model(ASSETS)
# Profile FastSAM-s
model = FastSAM("FastSAM-s.pt")
model.info()
model(ASSETS)
# Profile YOLO models (ONNX)
for file_name in ["yolov8n-seg.pt", "yolo11n-seg.pt", "yolo26n-seg.pt"]:
model = YOLO(file_name)
model.info()
onnx_path = model.export(format="onnx", dynamic=True, nms=False) # YOLO26 NMS-free head; no-op for YOLOv8/YOLO11
model = YOLO(onnx_path)
model(ASSETS)Auto-Annotation: A Quick Path to Segmentation Datasets#
Auto-annotation is a key feature of SAM, allowing users to generate a segmentation dataset using a pretrained detection model. This feature enables rapid annotation of large numbers of images, bypassing time-consuming manual labeling.
Generate Your Segmentation Dataset Using a Detection Model#
To auto-annotate your dataset with the Ultralytics framework, use the auto_annotate function, which pairs a detector such as YOLO26x with SAM:
from ultralytics.data.annotator import auto_annotate
auto_annotate(data="path/to/images", det_model="yolo26x.pt", sam_model="sam_b.pt")| Argument | Type | Default | Description |
|---|---|---|---|
data | str | required | Path to directory containing target images for annotation or segmentation. |
det_model | str | 'yolo26x.pt' | YOLO detection model path for initial object detection. |
sam_model | str | 'sam_b.pt' | SAM model path for segmentation (supports SAM, SAM 2, MobileSAM, and SAM 3 weights). |
device | str | '' | Computation device (e.g., 'cuda:0', 'cpu', or '' for automatic device detection). |
conf | float | 0.25 | YOLO detection confidence threshold for filtering weak detections. |
iou | float | 0.5 | IoU threshold for Non-Maximum Suppression to filter overlapping boxes. |
imgsz | int | 640 | Input size for resizing images (rounded up to a multiple of 32 if needed). |
max_det | int | 300 | Maximum number of detections per image for memory efficiency. |
classes | list[int] | None | List of class indices to detect (e.g., [0, 1] for person & bicycle). |
output_dir | str | None | Save directory for annotations (default: <data>_auto_annotate_labels sibling). |
The auto_annotate function takes the path to your images, with optional arguments for specifying the pretrained detection and SAM segmentation models, the device to run the models on, and the output directory for saving the annotated results.
Auto-annotation with pretrained models greatly reduces the time and effort required to create segmentation datasets, letting teams with large image collections focus on model development and evaluation rather than manual annotation.
Citations and Acknowledgments#
If you find SAM useful in your research or development work, please consider citing the paper:
@misc{kirillov2023segment,
title={Segment Anything},
author={Alexander Kirillov and Eric Mintun and Nikhila Ravi and Hanzi Mao and Chloe Rolland and Laura Gustafson and Tete Xiao and Spencer Whitehead and Alexander C. Berg and Wan-Yen Lo and Piotr Dollár and Ross Girshick},
year={2023},
eprint={2304.02643},
archivePrefix={arXiv},
primaryClass={cs.CV}
}We would like to express our gratitude to Meta AI for creating and maintaining this valuable resource for the computer vision community.
FAQ#
Meta's Segment Anything Model (SAM) is an image segmentation model designed for promptable segmentation tasks. It combines image and prompt encoders with a lightweight mask decoder to generate segmentation masks from prompts such as points, boxes, or masks. Trained on the SA-1B dataset, SAM performs well zero-shot, adapting to new image distributions and tasks without prior knowledge.
You can use the Segment Anything Model (SAM) for image segmentation by running inference with various prompts such as bounding boxes or points. Here's an example using Python:
from ultralytics import SAM # Load a model model = SAM("sam_b.pt") # Segment with bounding box prompt model("ultralytics/assets/zidane.jpg", bboxes=[439, 437, 524, 709]) # Segment with points prompt model("ultralytics/assets/zidane.jpg", points=[900, 370], labels=[1]) # Segment with multiple points prompt model("ultralytics/assets/zidane.jpg", points=[[400, 370], [900, 370]], labels=[1, 1]) # Segment with multiple points prompt per object model("ultralytics/assets/zidane.jpg", points=[[[400, 370], [900, 370]]], labels=[[1, 1]]) # Segment with negative points prompt. model("ultralytics/assets/zidane.jpg", points=[[[400, 370], [900, 370]]], labels=[[1, 0]])Alternatively, you can run inference with SAM in the command line interface (CLI):
yolo predict model=sam_b.pt source=path/to/image.jpgCompared to YOLO models, SAM variants like SAM-b, MobileSAM, and FastSAM-s are typically larger and slower but offer unique zero-shot segmentation capabilities. For example, YOLO26n-seg is 56x smaller and over 1650x faster than Meta's original SAM-b model on CPU. This makes YOLO models ideal for applications requiring rapid, lightweight, and computationally efficient segmentation, while SAM models excel in flexible, promptable, and zero-shot segmentation tasks.
Ultralytics' SAM offers an auto-annotation feature that allows generating segmentation datasets using a pretrained detection model. Here's an example in Python:
from ultralytics.data.annotator import auto_annotate auto_annotate(data="path/to/images", det_model="yolo26x.pt", sam_model="sam_b.pt")This function takes the path to your images and optional arguments for pretrained detection and SAM segmentation models, along with device and output directory specifications. For a complete guide, see Auto-Annotation.
SAM is trained on the SA-1B dataset, which comprises over 1 billion masks across 11 million images. SA-1B was the largest segmentation dataset at its release, providing high-quality and diverse training data that underpins SAM's zero-shot performance across segmentation tasks.
Comments27
Ask a question, share your results, or help someone else.
kh-diab
can we use SAM to process labeled images with bounding box to segment the object inside the box and generate new labels files like yolo or coco format if yes how please
1
Paula DerrengerUltralyticsYes, you can use SAM to process labeled images with bounding boxes to segment the objects inside and generate new label files in formats like YOLO or COCO. You can achieve this by using the
auto_annotatefunction in the Ultralytics framework. This function allows you to auto-annotate your dataset using a pre-trained detection model and SAM for segmentation. For more details, you can refer to the Segment Anything Model documentation. If you need further assistance, feel free to ask!kh-diab
Hello @pderrenger
Thanks I tried to run the below but I face error all the time
import os import cv2 import numpy as np from ultralytics import SAM
Load the SAM model
model = SAM("sam_b.pt") # Update the model path if needed
Function to load YOLO annotations
def load_yolo_annotations(annotation_path, image_width, image_height): bboxes = [] try: with open(annotation_path, 'r') as file: lines = file.readlines() for line in lines: parts = line.strip().split() class_id = int(parts[0]) x_center = float(parts[1]) * image_width y_center = float(parts[2]) * image_height width = float(parts[3]) * image_width height = float(parts[4]) * image_height x = x_center - width / 2 y = y_center - height / 2 bboxes.append((x, y, x + width, y + height)) # Update to [x1, y1, x2, y2] except Exception as e: print(f"Error reading annotation file {annotation_path}: {e}") return bboxes
Paths
images_folder = r'E:\MSc\tesi\20210901_Herrenpuent_Taenikon_Matrice_300_15m_1.6m_s\500 D' annotations_folder = r'C:\Users\Khaled\OneDrive - OST\Study Material\Msc Thesis\dataset\Group D yolo format\obj_train_data' output_folder = r'C:\Users\Khaled\OneDrive - OST\Study Material\Msc Thesis\dataset\test output sam'
if not os.path.exists(output_folder): os.makedirs(output_folder)
Process each image and its corresponding annotation
total_images = len(os.listdir(images_folder)) processed_images = 0
for image_filename in os.listdir(images_folder): if image_filename.endswith(('.JPG', '.jpg', '.png')): image_path = os.path.join(images_folder, image_filename) annotation_path = os.path.join(annotations_folder, image_filename.replace('.JPG', '.txt').replace('.jpg', '.txt').replace('.png', '.txt'))
print(f"Processing {image_filename}...") # Check if annotation file exists if not os.path.exists(annotation_path): print(f"Annotation file not found for {image_filename}") continue # Load the image image = cv2.imread(image_path) if image is None: print(f"Failed to load image {image_path}") continue image_rgb = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) height, width, _ = image.shape # Load YOLO annotations bboxes = load_yolo_annotations(annotation_path, width, height) if not bboxes: print(f"No bounding boxes found in {annotation_path}") continue # Perform segmentation using SAM model results = model(image_rgb, bboxes=bboxes) for idx, result in enumerate(results): if hasattr(result, 'masks') and result.masks is not None: masks = result.masks.cpu().numpy() # Convert to numpy array for m_idx, mask in enumerate(masks): chw_mask = mask.astype(np.uint8) # Ensure mask is in uint8 format hwc_mask = np.transpose(chw_mask, (1, 2, 0)) # Convert CHW to HWC contours, _ = cv2.findContours(hwc_mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE) segmented_image = cv2.bitwise_and(image, image, mask=hwc_mask) # Save the segmented image output_path = os.path.join(output_folder, image_filename.replace('.JPG', f'_segmented_{idx}_{m_idx}.jpg').replace('.jpg', f'_segmented_{idx}_{m_idx}.jpg').replace('.png', f'_segmented_{idx}_{m_idx}.png')) cv2.imwrite(output_path, segmented_image) print(f"Saved segmented image to {output_path}") processed_images += 1 print(f"Processed {processed_images}/{total_images} images.")print("Processing completed.")
error is
AttributeError: 'Masks' object has no attribute 'astype'. See valid attributes below.
A class for storing and manipulating detection masks. xy (list): A list of segments in pixel coordinates. xyn (list): A list of normalized segments. Methods: cpu(): Returns the masks tensor on CPU memory. numpy(): Returns the masks tensor as a numpy array. cuda(): Returns the masks tensor on GPU memory. to(device, dtype): Returns the masks tensor with the specified device and dtype.PS C:\Users\Khaled> & C:/MSc_Workspace/anaconda3/envs/GenEnv/python.exe "c:/Users/Khaled/OneDrive - OST/Study Material/Msc Thesis/coding/sam segment.py" Processing DJI_20210901122818_0695.JPG...
0: 1024x1024 67330.4ms Speed: 48.0ms preprocess, 67330.4ms inference, 8095.5ms postprocess per image at shape (1, 3, 1024, 1024) Traceback (most recent call last): File "c:\Users\Khaled\OneDrive - OST\Study Material\Msc Thesis\coding\sam segment.py", line 75, in chw_mask = mask.astype(np.uint8) # Ensure mask is in uint8 format File "C:\MSc_Workspace\anaconda3\envs\GenEnv\lib\site-packages\ultralytics\utils_init_.py", line 162, in getattr raise AttributeError(f"'{name}' object has no attribute '{attr}'. See valid attributes below.\n{self.doc}") AttributeError: 'Masks' object has no attribute 'astype'. See valid attributes below.
A class for storing and manipulating detection masks. Attributes: xy (list): A list of segments in pixel coordinates. xyn (list): A list of normalized segments. Methods: cpu(): Returns the masks tensor on CPU memory. numpy(): Returns the masks tensor as a numpy array. cuda(): Returns the masks tensor on GPU memory. to(device, dtype): Returns the masks tensor with the specified device and dtype.do you have idea how to solve it thanks
evans rojas
Hola me gustaría usar SAM para clasificar vegetación a partir de una imagen. ¿Qué recomendaciones me sugieren?
Paula DerrengerUltralyticsHola, Evans. SAM puede ayudarte a delimitar la vegetación, pero el SAM original no identifica por sí solo especies ni tipos de cobertura. Te sugiero comenzar con puntos positivos sobre la vegetación y negativos sobre el fondo, revisar las máscaras y asignarles tus clases manualmente. Si buscas automatizar la clasificación en más imágenes, puedes usar esas anotaciones para entrenar un modelo de segmentación por instancias para plantas individuales o de segmentación semántica para coberturas del terreno. ¿Quieres separar vegetación del fondo o distinguir especies, y trabajas con fotografías terrestres, de dron o satelitales?
meenakshi71
I've been trying to export the SAM model (which i had trained on my custom dataset) into ONNX but the output is always wrong. It is just generating a grid instead of a proper segmentation mask every time even after providing multiple kinds of prompts.
Glenn JocherUltralytics@meenakshi71 it seems the issue may stem from the export process or configuration. Ensure you are using the correct export parameters and supported formats. SAM models are primarily designed for real-time inference and not all functionalities may transfer seamlessly to ONNX. Refer to the SAM export documentation for guidance. If the problem persists, consider testing the ONNX model with simpler prompts to isolate the issue.
MeenakshiKK236
Why are my segmentation results being automatically saved into this directory while using the SAM model to perform segmentation on my custom dataset? "Results saved to runs/segment/predict4"
Glenn JocherUltralytics@MeenakshiKK236 the results are saved to the
runs/segment/predict4directory by default. You can change the save directory by specifying theprojectandnamearguments in your command. For more details, please refer to the SAM documentation.
NasserAlqeifi
Correct me if I'm wrong, when I use the auto_annotate function I will only get labels for my dataset with the classes in Yolo pre-trained? What if I need classes that are not available within Yolo? For example, I want to get a label for the road, and this is not available in any version of Yolo, but it is available in SAM?
Glenn JocherUltralytics@NasserAlqeifi you are correct. The
auto_annotatefunction uses YOLO's pre-trained classes for labeling. If you need classes not available in YOLO, like "road," you can use SAM directly for segmentation. For more details, please refer to the Segment Anything Model documentation.
Glenn JocherUltralytics@NasserAlqeifi but also remember SAM identifies segments but not classes, so it may find a sign or a road, but it won't be able to assign a class to them by itself.
NasserAlqeifi
Thank you for your quick response. Indeed, I tried to use SAM directly and found that it does not save the class names in the name, but the class number.
from ultralytics import SAM
Load a model
model = SAM("sam2_l.pt")
Predict results
results = model.predict(source=r"/testSAM/image.png", stream=True, bboxes=None, points=None, labels=None, show=False)
Iterate over results, print names, and save each result to a file
for i, result in enumerate(results): # Print the names detected in the result print(f"Result {i}: Detected Names - {result.names}")
# Save the result as an image file result.save(f"output_{i}.jpg")------------- output ---------
image 1/1 /testSAM/image.png: 1024x1024 1 0, 1 1, 1 2, 1 3, 1 4, 1 5, 1 6, 1 7, 1 8, 1 9, 1 10, 1 11, 1 12, 1 13, 1 14, 1 15, 1 16, 1 17, 1 18, 1 19, 1 20, 1 21, 1 22, 1 23, 1 24, 1 25, 1 26, 1 27, 1 28, 1 29, 1 30, 1 31, 1 32, 23758.2ms Result 0: Detected Names - {0: '0', 1: '1', 2: '2', 3: '3', 4: '4', 5: '5', 6: '6', 7: '7', 8: '8', 9: '9', 10: '10', 11: '11', 12: '12', 13: '13', 14: '14', 15: '15', 16: '16', 17: '17', 18: '18', 19: '19', 20: '20', 21: '21', 22: '22', 23: '23', 24: '24', 25: '25', 26: '26', 27: '27', 28: '28', 29: '29', 30: '30', 31: '31', 32: '32'} Speed: 20.3ms preprocess, 23758.2ms inference, 7.1ms postprocess per image at shape (1, 3, 1024, 1024)
Glenn JocherUltralyticsThank you for your feedback. Currently, SAM does not assign class names, only class numbers. If you need specific class names, you may need to implement a custom mapping or post-processing step to assign the desired class names based on your requirements.
dreambit
I have images where a single object (such as a car) is partially covered by small trees, resulting in the object being divided into multiple visible fragments. In coco it could look like { "id": 1, "image_id": 1, "category_id": 1, "bbox": [x, y, width, height], "segmentation": [ [polygon1_x1, polygon1_y1, polygon1_x2, polygon1_y2, ...], [polygon2_x1, polygon2_y1, polygon2_x2, polygon2_y2, ...] ], "area": 500, "iscrowd": 0 } Each polygon is an array of coordinates defining a closed shape
Is it possible with auto_annotate and yolo format?
Glenn JocherUltralytics(edited)@dreambit hi there!
Thank you for your detailed question. Yes, you can use the
auto_annotatefunction in the Ultralytics framework to handle images where objects are partially occluded and appear as multiple fragments. This function can generate segmentation masks for such complex scenarios.Here's how you can achieve this:
- Prepare Your Dataset: Ensure your images are organized in a directory.
- Use the
auto_annotateFunction: This function will use a pre-trained YOLO detection model and the SAM segmentation model to annotate your images.
Here's a code example to get you started:
from ultralytics.data.annotator import auto_annotate # Path to your images data_path = "path/to/your/images" # Auto-annotate using YOLO26 and SAM models auto_annotate(data=data_path, det_model="yolo26x.pt", sam_model="sam_b.pt")Explanation:
data: Path to the folder containing your images.det_model: The pre-trained YOLO detection model, which defaults toyolo26x.pt.sam_model: The pre-trained SAM segmentation model, which defaults tosam_b.pt.
This setup will generate segmentation masks for your images, even if the objects are partially occluded.
For more detailed information on using SAM and its capabilities, you can refer to the Segment Anything Model (SAM) documentation.
If you encounter any issues or need further assistance, please provide a reproducible example so we can better understand and address your specific scenario. You can find more information on creating a minimum reproducible example here.
Feel free to update to the latest versions of the packages to ensure you have the latest features and fixes.
Best of luck with your project! 🚀
dreambit
Glenn thanks, but as I see in docs https://docs.ultralytics.com/datasets/segment/#ultralytics-yolo-format ... In this format, is the index of the class for the object, and ... are the bounding coordinates of the object's segmentation mask. The coordinates are separated by spaces.
So it means that I could have only one segmentation mask?
dreambit
0 0.681 0.485 0.670 0.487 0.676 0.487 1 0.504 0.000 0.501 0.004 0.498 0.004 0.493 0.010 0.492 0.0104
dreambit
Glen, i ran the command you specified, and used an image from coco data set to test, as you can see segmentations are not correct, only half of the objects are segmented:
Here is original image from coco with car mask https://ibb.co/RjYyRdk
Glenn JocherUltralytics(edited)Hi @dreambit,
Thank you for sharing the images and details. I understand the challenge you're facing with partial segmentations. The
auto_annotatefunction should ideally handle such cases, but it seems like the segmentation results are not as expected.Steps to Troubleshoot and Improve Segmentation:
-
Verify Latest Versions: Ensure you are using the latest versions of the Ultralytics and SAM models. This can help avoid any bugs that might have been fixed in recent updates.
-
Reproducible Example: If the issue persists, providing a minimum reproducible example would be very helpful. This allows us to better understand and diagnose the problem. You can find guidance on creating a reproducible example here.
-
Adjusting Prompts: Sometimes, adjusting the prompts (bounding boxes or points) can improve segmentation results. Here’s an example of how you can use different prompts:
from ultralytics import SAM # Load a model model = SAM("sam_b.pt") # Run inference with bounding boxes prompt results = model("path/to/image.jpg", bboxes=[[x1, y1, x2, y2], [x3, y3, x4, y4]]) # Run inference with points prompt results = model("path/to/image.jpg", points=[[x1, y1], [x2, y2]], labels=[1, 1]) -
Segment Everything: If specific prompts are not yielding the desired results, you can try segmenting the entire image without prompts:
from ultralytics import SAM # Load a model model = SAM("sam_b.pt") # Segment the whole image results = model("path/to/image.jpg") -
Multiple Masks Handling: If your object is divided into multiple fragments, ensure that the segmentation model is capable of handling multiple masks per object. The YOLO format typically supports one segmentation mask per object, but you can adapt the output to handle multiple masks by combining them post-processing.
Example of Handling Multiple Masks:
If you need to combine multiple masks into a single annotation, you can do this post-processing:
# `results[0].masks.data` holds all masks of the image, shape (N, H, W) combined_mask = results[0].masks.data.cpu().numpy().max(axis=0) # Save or use the combined maskDocumentation and Resources:
For more detailed information on using SAM and its capabilities, please refer to the Segment Anything Model (SAM) documentation.
If you continue to face issues, please share a reproducible example, and we’ll be happy to assist further. Your feedback is valuable, and we aim to provide the best support possible.
-
kshweikeh
is there anyway where I can train YOLO SAM on a custom dataset?
Paula DerrengerUltralytics(edited)@kshweikeh hi there!
Currently, the Segment Anything Model (SAM) from Ultralytics is designed primarily for inference and does not support training on custom datasets. However, you can leverage SAM's powerful zero-shot capabilities for various segmentation tasks without the need for additional training.
If you're looking to create a segmentation dataset using SAM, you can use the auto-annotation feature. This allows you to generate segmentation masks using a pre-trained detection model. Here's a quick example:
from ultralytics.data.annotator import auto_annotate auto_annotate(data="path/to/images", det_model="yolo26x.pt", sam_model="sam_b.pt")This function will help you annotate your images quickly and efficiently, bypassing the need for manual labeling.
For more detailed information on using SAM, you can visit our Segment Anything Model (SAM) documentation.
If you have any further questions or need additional assistance, feel free to ask! 😊
Blessing988
@pderrenger It looks like auto_annotate for SAM only gives out the segmentation masks. What if you want to generate the bounding boxes in additionto the masks to verify if the results are right. Thanks.
Blessing988
Also, how will one visualise the segmentation masks after the auto_annotate which have been saved in a directory of txt files.
Paula DerrengerUltralytics(edited)Hi @Blessing988,
auto_annotatewrites one YOLO segmentation polygon per line (class_id x1 y1 x2 y2 ..., normalized) to its output folder (output_dir, by default<images folder>_auto_annotate_labels). Here is a script to draw those polygons on the images:from pathlib import Path import cv2 import numpy as np def visualize_annotations(image_dir, label_dir, output_dir): Path(output_dir).mkdir(parents=True, exist_ok=True) for image_path in Path(image_dir).glob("*.jpg"): image = cv2.imread(str(image_path)) h, w = image.shape[:2] label_path = Path(label_dir) / f"{image_path.stem}.txt" if not label_path.exists(): continue for line in label_path.read_text().splitlines(): values = line.split() polygon = (np.array(values[1:], dtype=float).reshape(-1, 2) * [w, h]).astype(np.int32) cv2.polylines(image, [polygon], isClosed=True, color=(0, 255, 0), thickness=2) cv2.imwrite(str(Path(output_dir) / image_path.name), image) visualize_annotations("path/to/images", "path/to/labels", "path/to/output")For more detailed information on using SAM and its features, you can refer to the Segment Anything Model (SAM) documentation.
Feel free to reach out if you have any further questions or need additional assistance! 😊
pabentse
Would be fantastic with support for the "H" size as well!
Paula DerrengerUltralytics@pabentse absolutely, the addition of an "H" size model would be a great enhancement to the range of model sizes we offer, catering to different computational needs and use cases. The YOLO community's feedback is invaluable, and suggestions like yours help guide future developments. Keep an eye on our updates! 👀🚀